# How Intuit Uses LLMs to Explain Taxes to Millions of Taxpayers

Jaspreet Singh, Intuit | AI Engineer World's Fair 2025 | 18:59

Source: https://www.youtube.com/watch?v=_zl_zimMRak
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/how-intuit-uses-llms-to-explain-taxes-to-millions-of-taxpayers
Published: 2025-07-23
Tags: evals, fine-tuning, guardrails, rag

## TL;DR
- Intuit uses Claude for tax explanations and OpenAI models for question answering in TurboTax.
- Tax experts create prompts and evaluation data, while automated LLM judges handle smaller iterations during development.
- The tax engine supplies tax numbers, and guardrails check generated explanations before users see them.

## Summary
Jaspreet Singh describes how Intuit built generative AI experiences for TurboTax, where users receive explanations of deductions, credits, refunds, and other tax outcomes. The system runs on Intuit's proprietary GenOS platform, with separate components for the user interface, orchestration, model selection, retrieval, and safety controls. Claude powers the main explanation workflow, while OpenAI models have been used for dynamic tax and product questions. Intuit also tested a fine-tuned Claude Haiku model to reduce prompt size and latency. Singh explains that tax experts write prompts, establish manual evaluations, and create the data used to train automated LLM judges. The tax knowledge engine remains responsible for calculations. LLMs explain those results, while safety models check for hallucinated numbers. The talk is candid about vendor lock-in, changing models, expensive evaluations, and latency spikes near tax deadlines. Intuit uses RAG and graphRAG for tax questions, with personalization improving the usefulness of answers.

## Key ideas
### TurboTax uses explanations to help users understand their tax outcomes
[00:03](https://www.youtube.com/watch?v=_zl_zimMRak&t=3s)
Intuit processed 44 million tax returns for tax year 2023, and Singh says the product needs to give users confidence in how their taxes are filed. TurboTax already knows the user's deductions, credits, standard deduction, W-2 withholding, and other components. The generative experience expands on those inputs and explains how they affect the user's tax breaks and overall refund. The goal is to make the result understandable rather than simply showing a final number.

### Intuit built its own GenOS because generic tooling did not fit regulated tax work
[01:31](https://www.youtube.com/watch?v=_zl_zimMRak&t=91s)
Intuit's generative AI experiences run on GenOS, an internal platform with a user interface layer, an orchestrator, and components for different teams and model solutions. Singh says off-the-shelf generative AI tooling did not support all of Intuit's use cases. Tax work requires strong attention to safety and security because Intuit operates in a regulated business. The platform also has to support end-to-end use across a large company, rather than a single experimental application.

### Different tax questions use different prompts, models, and retrieval methods
[02:57](https://www.youtube.com/watch?v=_zl_zimMRak&t=177s)
The first version of the explanation system used prompts to describe the components of a user's refund. Static queries are prepared for known parts of the experience, such as a summary page. Dynamic queries answer questions about a user's particular situation, such as whether a deduction is available. Claude became the production model for the explanation use case, while OpenAI models were used for other question answering. Because IRS forms and rules change, Intuit also uses retrieval-augmented generation and graphRAG with its tax information.

### Fine-tuning Claude Haiku can reduce instructions and prompt length
[07:20](https://www.youtube.com/watch?v=_zl_zimMRak&t=440s)
Intuit piloted a fine-tuned version of Claude 3 Haiku through AWS Bedrock for static tax queries. Singh says the model quality was good, but fine-tuning required effort and produced a model that was specialized for the particular use case. The main reasons to try it were to reduce the number of instructions required and shrink the prompt. That could help with latency while preserving the quality needed for tax explanations. Intuit trained only on consented user data and used separate AWS accounts and environments.

### Tax experts provide the starting point for prompts and evaluations
[05:36](https://www.youtube.com/watch?v=_zl_zimMRak&t=336s)
Intuit works with tax analysts who interpret yearly IRS changes and provide domain knowledge. They perform the first manual evaluations and also act as prompt engineers. This lets data science and machine learning teams focus on defining quality metrics, building datasets, and testing models. The manual expert work becomes the basis for automated evaluations. Singh describes accuracy, relevancy, and coherence as the main evaluation dimensions.

### Automated LLM judges handle small changes after experts establish a baseline
[08:45](https://www.youtube.com/watch?v=_zl_zimMRak&t=525s)
The evaluation process changes as development progresses. Early work uses manual reviews by tax experts to establish a baseline. Once that baseline exists, smaller prompt changes can use automated evaluation. Intuit builds an LLM judge from expert-labelled examples and a golden dataset, with AWS Ground Truth used for the manual samples. Larger changes still require human review. Singh gives the move from tax year 2023 to tax year 2024 as an example of a change that needs renewed manual evaluation.

### The tax engine calculates numbers, while LLMs explain them
[15:15](https://www.youtube.com/watch?v=_zl_zimMRak&t=915s)
TurboTax has a proprietary tax knowledge engine that supplies the numerical tax results. The LLM does not calculate the user's taxes. It receives the existing results and formulates an explanation around them. Intuit also has safety guardrails and machine learning checks that inspect the raw response for hallucinated numbers before it reaches the user. This separation lets the established tax system remain the source of numerical truth while generative AI handles explanation.

### Latency and vendor lock-in shape the product design
[10:18](https://www.youtube.com/watch?v=_zl_zimMRak&t=618s)
Singh says model contracts can cost millions of dollars and long-term contracts can tie a company to a vendor. Prompts also create lock-in, since changing models or even upgrading within the same vendor can require substantial work. Latency is another practical problem. LLM responses can take three, five, or ten seconds rather than the response times expected from ordinary backend services. A complicated household tax situation creates larger prompts, and latency can rise sharply around April 15. Intuit therefore designs fallbacks and user-facing flows around these limits.

### Personalized graphRAG answers are more useful than generic retrieval
[16:41](https://www.youtube.com/watch?v=_zl_zimMRak&t=1001s)
Intuit uses both traditional RAG and graphRAG for tax question answering. Singh says graphRAG has produced better response quality than regular RAG in their testing. Personalization matters even more for end-user helpfulness, because the answer needs to match the user's own tax situation. The question-answering system covers product questions, such as how to perform an action in TurboTax, as well as tax questions about situations such as claiming tuition paid for a grandchild. A planner routes different questions to different solutions.

## Notable quotes
- "Evals are a must to launch. Focus on evals. Make sure you have clear guidelines on what you're building." (12:00)
- "We are not having LLMs do the calculations at all." (15:55)
- "Vendors are a form of lock in. The prompts are a form of lock in." (10:40)
- "For end user helpfulness getting personalized answer is the key piece." (17:14)

## Tools & references mentioned
- Intuit
- TurboTax
- GenOS
- GenUX
- Intuit Assist
- Anthropic
- Claude
- Claude 3 Haiku
- Claude 4
- OpenAI
- GPT-4o mini
- AWS Bedrock
- AWS Ground Truth
- IRS
- RAG
- graphRAG

## Who should watch
- You are shipping generative AI in a regulated product and need a concrete evaluation process with domain experts.
- Your application must explain outputs from an existing rules or calculation engine without letting an LLM invent the underlying numbers.
- You are weighing model fine-tuning, retrieval, vendor contracts, and latency for a high-volume production system.

## Related talks

- [Why Off-the-Shelf AI Doesn't Understand Money](https://aietalks.com/talks/why-off-the-shelf-ai-doesnt-understand-money) (Udi Menkes, Intuit, 19:50)
- [Training Albatross, An Expert Finance LLM](https://aietalks.com/talks/training-albatross-an-expert-finance-llm) (Leo Pekelis, Gradient, 16:20)
- [Real ROI: Lessons from Enterprises That Have Already Succeeded with LLMs at Scale](https://aietalks.com/talks/real-roi-lessons-from-enterprises-that-have-already-succeeded-with-llms-at-scale) (Raza Habib, Humanloop, 20:01)
- [Transforming search and discovery using LLMs](https://aietalks.com/talks/transforming-search-and-discovery-using-llms) (Tejaswi Tenneti & Vinesh Gudla, Instacart, 21:10)
- [Low Level Technicals of LLMs](https://aietalks.com/talks/low-level-technicals-of-llms) (Daniel Han, Unsloth, 2:52:26)